Prediction and mapping of soil organic carbon in the Bosten Lake oasis based on Sentinel-2 data and environmental variables

• Published in: International Soil and Water Conservation Research Vol. 13; no. 2; pp. 436 - 446
• Main Authors: Li, Shaotian, Li, Xinguo, Ge, Xiangyu
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.06.2025; KeAi Communications Co., Ltd
• Subjects: agricultural land; carbon sinks; China; color; Google earth engine; lakes; land use; Machine learning; oases; prediction; Sentinel-2A; soil; Soil organic carbon; Spectral index; support vector machines; vegetation index; water conservation; watersheds; woodlands; China; Spectral index; Sentinel-2A; Google earth engine; Soil organic carbon; Machine learning
• ISSN: 2095-6339
• DOI: 10.1016/j.iswcr.2024.12.002

Soil is the largest carbon pool on the Earth's surface. With the application of remote sensing technology, Soil Organic Carbon (SOC) estimation has become a hot topic in digital soil mapping. However, the heterogeneity of geomorphology can affect the performance of remote sensing in determining soil organic carbon. In the Bosten Lake Watershed in northwestern China, we collected 116 soil samples from farm land, uncultivated land, and woodland. To establish an SOC prediction model, we produced 16 optical remote sensing variables and 9 environmental covariates. Three types of land use were studied: farm land, uncultivated land, and woodland. Five machine learning models were used for these land use types: gradient Tree (ET), Support Vector Machine (SVM), Random Forest (RF), Adaptive gradient Boosting (AdaBoost), and extreme Gradient Boosting (XGBoost). The main driving variables for changes in organic carbon content across the entire sample area were Enhanced Vegetation Index (EVI), Enhanced Vegetation Index 2 (EVI2), Soil-Adjusted Vegetation Index (SAVI); for farm land, it was Clay Index (CI2); for farm land and woodland, it was Color Index (CI). The results showed that in terms of prediction accuracy, RF and XGBoost outperformed SVM. In terms of simulation precision, the ET model's woodland model (R2 = 0.86, RMSE = 7.72), the ET model's farm land model (R2 = 0.82, RMSE = 6.66), and the uncultivated land model of the RF model (R2 = 0.81, RMSE = 1.09) performed best. Compared to global modeling, establishing SOC estimation models based on different land use types yielded more ideal results in this study. These findings provide new insights into high-precision estimation of organic carbon content.